Screen Backtest

Two layers of evidence. Company-level (2018–2026): for each month-end we rebuilt what each screen would have shown using only SEC filings submitted by that date, then measured whether each pick beat the median $250M+ company over the following 12 months (a coin flip = 50%). Decades-level (1926–2026, below): each screen's academic style cousin — survivorship-free portfolios sorted on the same characteristic — scored across every economic regime since 1955, including the real stagflation of the 1970s that the modern sample lacks. Structured SEC data only exists from ~2009 and our price history from 2016, which is why company-level testing cannot reach further back.

Read this before quoting numbers: the test universe is our 2026 SEC-data coverage, so companies that delisted before 2026 are missing — a survivorship bias every screen shares equally. That makes screen-vs-screen and screen-vs-universe comparisons fair, but flatters absolute numbers (distress-flavored screens most). Also: 2018–2026 was a mega-cap index decade — even the best screen only ties SPY, while clearly beating the median stock. Screens with few observations are anecdotes, not evidence. Full methodology: experiments/012_preset_backtest/README.md.

Single screens

Share of picks beating the same-month universe median over the following 12 months. H1 = 2018 to mid-2022, H2 = mid-2022 on.

Month by month — is a screen's edge steady, or one lucky stretch? Each point: the share of that month's picks that beat the universe median over the FOLLOWING 12 months. A real edge hugs the upper side of the 50% line for years at a time.

Screen

One to three years after FIRST hitting a signal

The event-study view: instead of counting every month a company sits inside a screen, this counts only the ENTRY month — the first time a signal turns true after three observed months of being false — and follows each company 12, 18, 24 and 36 months out. Cells show % beating the universe median / median excess return. Two lessons jump out: value-type signals BUILD for two to three years, and the start of a heavy building phase is a multi-year drag.

Structural events (8-K filings). Contracts, acquisitions, leadership changes and other material events are disclosed on Form 8-K with machine-readable item codes — scored here the same way: first filing of each type in 3+ months, followed 12/24/36 months against the universe median.

Slice one signal by sub-industry (NASDAQ granularity — copper & precious metals, auto parts, software…). Only industries with 8+ events shown; 24-month horizon.

Signal

Combinations

All 74 pairs tested. proven = beat BOTH parent screens over the full period AND in each half — the guard against crowning the luckiest of many tries. Click a pairing to load it in the screener.

Sectors by economic regime

Average 3-month sector return minus the universe, per regime (equal-weight across our $250M+ coverage). Note: no stagflation months (inflation ≥4% with weakening jobs) occurred in 2018–2026.

The same table at sub-industry granularity (106 industries with enough data; needs 3+ companies per month and 8+ months per regime — thin cells excluded). Scroll inside the box.

Across the decades — each screen's style cousin, 1926–2026

Company-level screening data cannot reach past ~2016, so this layer maps each screen to its closest academic style (Ken French research portfolios: value-weighted, survivorship-free — delisted companies included). Cousins, not clones: they sort on one characteristic where our screens demand several. Numbers are the style portfolio's return minus the market, annualized.

Combinations across the decades. French also publishes DOUBLE-sorted portfolios (two characteristics at once, same survivorship-free construction, 1963+) — combination tests at 60-year scale. Only these pairings exist academically; arbitrary pairs would need the underlying licensed data. The two "double traps" are the anti-screens.

Buy today, hold 1–3 years — a century of entry points. The decades version of the event study: from EVERY month-end since 1926, the style's cumulative return vs the market over the following 12 / 24 / 36 months. Cells: % of windows ahead / median excess over the window. Right half: only windows STARTING inside stagflation. The same patience law as the modern event study — real edges build with horizon, traps compound.

Decade by decade — did the style work in most eras, or just one?

Industries by regime since 1926 (French 10-industry portfolios) — the long-history version of the sector table above.

The century in one chart. Each line: growth of $1 in the style divided by growth of $1 in the market (log scale) — rising means the style is beating the market, flat means no edge, falling means it lags. Gold bands are the stagflation stretches. This is the raw arithmetic behind the tables above: every claimed edge should be visible as a long grind upward, and every "lost decade" is visible too (watch cheap-on-assets go sideways through the 2010s). Click legend entries to hide or show lines.

Time machine

Pick a screen and a month: the exact companies it listed that day, and what each did next. This is the raw evidence behind every number above.

Screen at month-end